Tidemark

Inspiration

What it does

How we built it

Inspiration

Every memory system on the market is competing on how much it can remember. We built the opposite bet: an agent memory that forgets on purpose, learns only from outcomes, and can prove every recall it ever made.

Remembering something is not the same as it being useful. Tidemark only strengthens a memory when a task outcome — with item-level evidence — credits it. Memories nobody's outcomes ever vouch for simply fade, on a half-life. And every recall opens a receipt: rank, similarity, strength, utility, all persisted in CockroachDB, all auditable after the fact.

What it does

• Outcome-gated plasticity: recall alone changes nothing (read-only, receipt-only). Only a terminal outcome with item-bound evidence moves a memory's long-term anchor — credited memories consolidate, blamed ones sink, and a failed task never punishes a memory without evidence.
• Honest forgetting: read-time decay on a half-life; pinned memories freeze; nothing is silently deleted — owner deletion propagates with tombstones.
• Memory receipts: every recall persists its score components. The Evidence console shows the receipt, the plasticity before/after, and the server-sampled decay curve for any record — nothing is reconstructed client-side.
• A judge-runnable proof: one button runs a ten-step lifecycle proof on the real production path (real tools, real vector recall, real CockroachDB writes) and asserts the three claims that matter: recall mutates nothing, a terminal outcome credits only the memory with evidence, and the state survives a fresh transaction.

How we built it

• CockroachDB (tool 1): distributed vector indexing over VECTOR(512) with an embedding-identity index prefix — two embedding spaces can never be searched together. All memories, receipts, evidence ledgers, and nightly runs live here in SERIALIZABLE transactions.
• CockroachDB Cloud Managed MCP Server (tool 2): an operator-facing read-only audit path over sanitized views — evidenced live on 2026-08-12 (docs/EVIDENCE-MANAGED-MCP-0812.md), not just configured.
• AWS: Lambda runs the Memory MCP server and the nightly orchestrator — including in-process ONNX embedding (sealed, SHA256-pinned MiniLM; no external AI API in the request path). API Gateway fronts it; EventBridge schedules the nightly dream/reflection/transition pipeline with idempotent, lease-fenced runs and an SQS DLQ; Secrets Manager holds the only secret.
• Zero-credential demo: CloudFront serves the console from a private S3 origin and injects the read-only viz key as an origin custom header on the way to API Gateway. The browser holds nothing; the viz key cannot reach the tool face.

Challenges we ran into

• Bedrock access was formally denied for this individual account, so we moved embedding inside the Lambda — sealed ONNX artifacts, manifest-verified at cold start, bit-exact across dev and Lambda.
• CN-route network storms (ECONNRESET mid-BEGIN) taught the connection pool to evict dirty connections that came back holding a server-side transaction.
• Honesty is a feature with a maintenance cost: every capability the console shows carries live / documented / evidence_pending / blocked_external, with a reason and a repo pointer — and tests that fail if a pending item ever claims to be live.

Accomplishments that we're proud of

• A 21-round evaluation loop (three-arm A/B: no-memory / vector-only / full) showing injection-hit and lifecycle ablation on deterministic scripted agents — main success 0 / 0.875 / 1.0.
• A natural-decay E2E: after 2.3 real days, served values matched the decay formula with zero per-row delta, and credited memories out-survived their counterfactual.
• Roughly 100 cross-review conclusions between two AI engineers (Claude x Codex) logged in collab/CODEX_CHANNEL.md — the decision log ships with the repo.

What's next

Contradiction links (candidate edges with evidence, never auto-suppression), the model segments of dream/reflection when Bedrock access lands, and abstain calibration for the A/B harness.

Team

• Ovo — owner, direction, product rulings, operator evidence
• Claude (Anthropic) — architecture, backend, evidence frontend, docs
• Codex (OpenAI) — second-perspective engineering, cross-review, frontend build ## Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Tidemark

Built With

• api-gateway
• aws-lambda
• cloudfront
• cockroachdb
• eventbridge
• mcp
• node.js
• onnx
• react
• s3
• secrets-manager
• sqs
• transformers.js
• typescript